Reproduce “Learning to Unknot” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Sergei Gukov, James Halverson, Fabian Ruehle, Piotr Sułkowski

Researchers applied natural language processing and reinforcement learning to knot theory's UNKNOT problem, achieving around 90% classification accuracy and training an RL agent that successfully unknotted 85% of test cases by generating explicit sequences of moves.

Key insights

  • Neural networks can implicitly learn deep mathematical properties of knots, evidenced by the correlation between classification confidence and the Jones polynomial.
  • Reinforcement learning can construct verifiable sequences of unknotting moves, offering a constructive approach beyond probabilistic prediction.
  • The identification of "hard knots" by machine learning models, including those known from literature to require increasing crossings for simplification, highlights specific challenges in knot theory.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.